net/rds/tcp.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35 #include <linux/in.h>
  36 #include <linux/module.h>
  37 #include <net/tcp.h>
  38 #include <net/net_namespace.h>
  39 #include <net/netns/generic.h>
  40 #include <net/tcp.h>
  41
  42 #include "rds.h"
  43 #include "tcp.h"
  44
  45 /* only for info exporting */
  46 static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
  47 static LIST_HEAD(rds_tcp_tc_list);
  48 static unsigned int rds_tcp_tc_count;
  49
  50 /* Track rds_tcp_connection structs so they can be cleaned up */
  51 static DEFINE_SPINLOCK(rds_tcp_conn_lock);
  52 static LIST_HEAD(rds_tcp_conn_list);
  53
  54 static struct kmem_cache *rds_tcp_conn_slab;
  55
  56 #define RDS_TCP_DEFAULT_BUFSIZE (128 * 1024)
  57
  58 /* doing it this way avoids calling tcp_sk() */
  59 void rds_tcp_nonagle(struct socket *sock)
  60 {
  61         mm_segment_t oldfs = get_fs();
  62         int val = 1;
  63
  64         set_fs(KERNEL_DS);
  65         sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val,
  66                               sizeof(val));
  67         set_fs(oldfs);
  68 }
  69
  70 /* All module specific customizations to the RDS-TCP socket should be done in
  71  * rds_tcp_tune() and applied after socket creation. In general these
  72  * customizations should be tunable via module_param()
  73  */
  74 void rds_tcp_tune(struct socket *sock)
  75 {
  76         rds_tcp_nonagle(sock);
  77 }
  78
  79 u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
  80 {
  81         return tcp_sk(tc->t_sock->sk)->snd_nxt;
  82 }
  83
  84 u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
  85 {
  86         return tcp_sk(tc->t_sock->sk)->snd_una;
  87 }
  88
  89 void rds_tcp_restore_callbacks(struct socket *sock,
  90                                struct rds_tcp_connection *tc)
  91 {
  92         rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
  93         write_lock_bh(&sock->sk->sk_callback_lock);
  94
  95         /* done under the callback_lock to serialize with write_space */
  96         spin_lock(&rds_tcp_tc_list_lock);
  97         list_del_init(&tc->t_list_item);
  98         rds_tcp_tc_count--;
  99         spin_unlock(&rds_tcp_tc_list_lock);
 100
 101         tc->t_sock = NULL;
 102
 103         sock->sk->sk_write_space = tc->t_orig_write_space;
 104         sock->sk->sk_data_ready = tc->t_orig_data_ready;
 105         sock->sk->sk_state_change = tc->t_orig_state_change;
 106         sock->sk->sk_user_data = NULL;
 107
 108         write_unlock_bh(&sock->sk->sk_callback_lock);
 109 }
 110
 111 /*
 112  * This is the only path that sets tc->t_sock.  Send and receive trust that
 113  * it is set.  The RDS_CONN_CONNECTED bit protects those paths from being
 114  * called while it isn't set.
 115  */
 116 void rds_tcp_set_callbacks(struct socket *sock, struct rds_connection *conn)
 117 {
 118         struct rds_tcp_connection *tc = conn->c_transport_data;
 119
 120         rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
 121         write_lock_bh(&sock->sk->sk_callback_lock);
 122
 123         /* done under the callback_lock to serialize with write_space */
 124         spin_lock(&rds_tcp_tc_list_lock);
 125         list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
 126         rds_tcp_tc_count++;
 127         spin_unlock(&rds_tcp_tc_list_lock);
 128
 129         /* accepted sockets need our listen data ready undone */
 130         if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
 131                 sock->sk->sk_data_ready = sock->sk->sk_user_data;
 132
 133         tc->t_sock = sock;
 134         tc->conn = conn;
 135         tc->t_orig_data_ready = sock->sk->sk_data_ready;
 136         tc->t_orig_write_space = sock->sk->sk_write_space;
 137         tc->t_orig_state_change = sock->sk->sk_state_change;
 138
 139         sock->sk->sk_user_data = conn;
 140         sock->sk->sk_data_ready = rds_tcp_data_ready;
 141         sock->sk->sk_write_space = rds_tcp_write_space;
 142         sock->sk->sk_state_change = rds_tcp_state_change;
 143
 144         write_unlock_bh(&sock->sk->sk_callback_lock);
 145 }
 146
 147 static void rds_tcp_tc_info(struct socket *sock, unsigned int len,
 148                             struct rds_info_iterator *iter,
 149                             struct rds_info_lengths *lens)
 150 {
 151         struct rds_info_tcp_socket tsinfo;
 152         struct rds_tcp_connection *tc;
 153         unsigned long flags;
 154         struct sockaddr_in sin;
 155         int sinlen;
 156
 157         spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
 158
 159         if (len / sizeof(tsinfo) < rds_tcp_tc_count)
 160                 goto out;
 161
 162         list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
 163
 164                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 0);
 165                 tsinfo.local_addr = sin.sin_addr.s_addr;
 166                 tsinfo.local_port = sin.sin_port;
 167                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 1);
 168                 tsinfo.peer_addr = sin.sin_addr.s_addr;
 169                 tsinfo.peer_port = sin.sin_port;
 170
 171                 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
 172                 tsinfo.data_rem = tc->t_tinc_data_rem;
 173                 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
 174                 tsinfo.last_expected_una = tc->t_last_expected_una;
 175                 tsinfo.last_seen_una = tc->t_last_seen_una;
 176
 177                 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
 178         }
 179
 180 out:
 181         lens->nr = rds_tcp_tc_count;
 182         lens->each = sizeof(tsinfo);
 183
 184         spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
 185 }
 186
 187 static int rds_tcp_laddr_check(struct net *net, __be32 addr)
 188 {
 189         if (inet_addr_type(net, addr) == RTN_LOCAL)
 190                 return 0;
 191         return -EADDRNOTAVAIL;
 192 }
 193
 194 static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 195 {
 196         struct rds_tcp_connection *tc;
 197
 198         tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
 199         if (!tc)
 200                 return -ENOMEM;
 201
 202         tc->t_sock = NULL;
 203         tc->t_tinc = NULL;
 204         tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 205         tc->t_tinc_data_rem = 0;
 206
 207         conn->c_transport_data = tc;
 208
 209         spin_lock_irq(&rds_tcp_conn_lock);
 210         list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
 211         spin_unlock_irq(&rds_tcp_conn_lock);
 212
 213         rdsdebug("alloced tc %p\n", conn->c_transport_data);
 214         return 0;
 215 }
 216
 217 static void rds_tcp_conn_free(void *arg)
 218 {
 219         struct rds_tcp_connection *tc = arg;
 220         unsigned long flags;
 221         rdsdebug("freeing tc %p\n", tc);
 222
 223         spin_lock_irqsave(&rds_tcp_conn_lock, flags);
 224         list_del(&tc->t_tcp_node);
 225         spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
 226
 227         kmem_cache_free(rds_tcp_conn_slab, tc);
 228 }
 229
 230 static void rds_tcp_destroy_conns(void)
 231 {
 232         struct rds_tcp_connection *tc, *_tc;
 233         LIST_HEAD(tmp_list);
 234
 235         /* avoid calling conn_destroy with irqs off */
 236         spin_lock_irq(&rds_tcp_conn_lock);
 237         list_splice(&rds_tcp_conn_list, &tmp_list);
 238         INIT_LIST_HEAD(&rds_tcp_conn_list);
 239         spin_unlock_irq(&rds_tcp_conn_lock);
 240
 241         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 242                 if (tc->conn->c_passive)
 243                         rds_conn_destroy(tc->conn->c_passive);
 244                 rds_conn_destroy(tc->conn);
 245         }
 246 }
 247
 248 static void rds_tcp_exit(void);
 249
 250 struct rds_transport rds_tcp_transport = {
 251         .laddr_check            = rds_tcp_laddr_check,
 252         .xmit_prepare           = rds_tcp_xmit_prepare,
 253         .xmit_complete          = rds_tcp_xmit_complete,
 254         .xmit                   = rds_tcp_xmit,
 255         .recv                   = rds_tcp_recv,
 256         .conn_alloc             = rds_tcp_conn_alloc,
 257         .conn_free              = rds_tcp_conn_free,
 258         .conn_connect           = rds_tcp_conn_connect,
 259         .conn_shutdown          = rds_tcp_conn_shutdown,
 260         .inc_copy_to_user       = rds_tcp_inc_copy_to_user,
 261         .inc_free               = rds_tcp_inc_free,
 262         .stats_info_copy        = rds_tcp_stats_info_copy,
 263         .exit                   = rds_tcp_exit,
 264         .t_owner                = THIS_MODULE,
 265         .t_name                 = "tcp",
 266         .t_type                 = RDS_TRANS_TCP,
 267         .t_prefer_loopback      = 1,
 268 };
 269
 270 static int rds_tcp_netid;
 271
 272 /* per-network namespace private data for this module */
 273 struct rds_tcp_net {
 274         struct socket *rds_tcp_listen_sock;
 275         struct work_struct rds_tcp_accept_w;
 276 };
 277
 278 static void rds_tcp_accept_worker(struct work_struct *work)
 279 {
 280         struct rds_tcp_net *rtn = container_of(work,
 281                                                struct rds_tcp_net,
 282                                                rds_tcp_accept_w);
 283
 284         while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
 285                 cond_resched();
 286 }
 287
 288 void rds_tcp_accept_work(struct sock *sk)
 289 {
 290         struct net *net = sock_net(sk);
 291         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 292
 293         queue_work(rds_wq, &rtn->rds_tcp_accept_w);
 294 }
 295
 296 static __net_init int rds_tcp_init_net(struct net *net)
 297 {
 298         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 299
 300         rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
 301         if (!rtn->rds_tcp_listen_sock) {
 302                 pr_warn("could not set up listen sock\n");
 303                 return -EAFNOSUPPORT;
 304         }
 305         INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
 306         return 0;
 307 }
 308
 309 static void __net_exit rds_tcp_exit_net(struct net *net)
 310 {
 311         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 312
 313         /* If rds_tcp_exit_net() is called as a result of netns deletion,
 314          * the rds_tcp_kill_sock() device notifier would already have cleaned
 315          * up the listen socket, thus there is no work to do in this function.
 316          *
 317          * If rds_tcp_exit_net() is called as a result of module unload,
 318          * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
 319          * we do need to clean up the listen socket here.
 320          */
 321         if (rtn->rds_tcp_listen_sock) {
 322                 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 323                 rtn->rds_tcp_listen_sock = NULL;
 324                 flush_work(&rtn->rds_tcp_accept_w);
 325         }
 326 }
 327
 328 static struct pernet_operations rds_tcp_net_ops = {
 329         .init = rds_tcp_init_net,
 330         .exit = rds_tcp_exit_net,
 331         .id = &rds_tcp_netid,
 332         .size = sizeof(struct rds_tcp_net),
 333 };
 334
 335 static void rds_tcp_kill_sock(struct net *net)
 336 {
 337         struct rds_tcp_connection *tc, *_tc;
 338         struct sock *sk;
 339         LIST_HEAD(tmp_list);
 340         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 341
 342         rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 343         rtn->rds_tcp_listen_sock = NULL;
 344         flush_work(&rtn->rds_tcp_accept_w);
 345         spin_lock_irq(&rds_tcp_conn_lock);
 346         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 347                 struct net *c_net = read_pnet(&tc->conn->c_net);
 348
 349                 if (net != c_net || !tc->t_sock)
 350                         continue;
 351                 list_move_tail(&tc->t_tcp_node, &tmp_list);
 352         }
 353         spin_unlock_irq(&rds_tcp_conn_lock);
 354         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 355                 sk = tc->t_sock->sk;
 356                 sk->sk_prot->disconnect(sk, 0);
 357                 tcp_done(sk);
 358                 if (tc->conn->c_passive)
 359                         rds_conn_destroy(tc->conn->c_passive);
 360                 rds_conn_destroy(tc->conn);
 361         }
 362 }
 363
 364 static int rds_tcp_dev_event(struct notifier_block *this,
 365                              unsigned long event, void *ptr)
 366 {
 367         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 368
 369         /* rds-tcp registers as a pernet subys, so the ->exit will only
 370          * get invoked after network acitivity has quiesced. We need to
 371          * clean up all sockets  to quiesce network activity, and use
 372          * the unregistration of the per-net loopback device as a trigger
 373          * to start that cleanup.
 374          */
 375         if (event == NETDEV_UNREGISTER_FINAL &&
 376             dev->ifindex == LOOPBACK_IFINDEX)
 377                 rds_tcp_kill_sock(dev_net(dev));
 378
 379         return NOTIFY_DONE;
 380 }
 381
 382 static struct notifier_block rds_tcp_dev_notifier = {
 383         .notifier_call        = rds_tcp_dev_event,
 384         .priority = -10, /* must be called after other network notifiers */
 385 };
 386
 387 static void rds_tcp_exit(void)
 388 {
 389         rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 390         unregister_pernet_subsys(&rds_tcp_net_ops);
 391         if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
 392                 pr_warn("could not unregister rds_tcp_dev_notifier\n");
 393         rds_tcp_destroy_conns();
 394         rds_trans_unregister(&rds_tcp_transport);
 395         rds_tcp_recv_exit();
 396         kmem_cache_destroy(rds_tcp_conn_slab);
 397 }
 398 module_exit(rds_tcp_exit);
 399
 400 static int rds_tcp_init(void)
 401 {
 402         int ret;
 403
 404         rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
 405                                               sizeof(struct rds_tcp_connection),
 406                                               0, 0, NULL);
 407         if (!rds_tcp_conn_slab) {
 408                 ret = -ENOMEM;
 409                 goto out;
 410         }
 411
 412         ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
 413         if (ret) {
 414                 pr_warn("could not register rds_tcp_dev_notifier\n");
 415                 goto out;
 416         }
 417
 418         ret = register_pernet_subsys(&rds_tcp_net_ops);
 419         if (ret)
 420                 goto out_slab;
 421
 422         ret = rds_tcp_recv_init();
 423         if (ret)
 424                 goto out_slab;
 425
 426         ret = rds_trans_register(&rds_tcp_transport);
 427         if (ret)
 428                 goto out_recv;
 429
 430         rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 431
 432         goto out;
 433
 434 out_recv:
 435         rds_tcp_recv_exit();
 436 out_slab:
 437         unregister_pernet_subsys(&rds_tcp_net_ops);
 438         kmem_cache_destroy(rds_tcp_conn_slab);
 439 out:
 440         return ret;
 441 }
 442 module_init(rds_tcp_init);
 443
 444 MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
 445 MODULE_DESCRIPTION("RDS: TCP transport");
 446 MODULE_LICENSE("Dual BSD/GPL");
 447